human metacognition inspired learning algorithms in neural

Human Metacognition Inspired Learning Algorithms in Neural Networks and Particle

Swarm Optimization

Asst. Prof. Sundaram Suresh Nanyang Technological University, Singapore

Email : [email protected]

Webpage: www.ntu.edu.sg/home/ssundaram

IEEE SSCI 2015 Tutorial

Cape Town, South Africa

S. Suresh, SCE, NTU 8/12/2015 1

mailto:[email protected]

Organization • A brief review on human learning from cognitive psychology. • Part I – Neural Networks

– Self-regulated learning with meta-cognition in machine learning – Meta-cognitive neural networks and its self-regulatory learning algorithms

• PBL-McRBFN classifier – Benchmark evaluation and comparisons – Applications in Medical informatics

• Part II – Particle Swarm Optimization – Self-regulated particle swarm optimization – Dynamic mentoring based particle swarm optimization – Real-world applications

• Future directions

8/12/2015 S. Suresh, SCE, NTU 2

Motivation • Self-regulated learning is the best

learning strategy [1],[2],[3]

• Student control their learning process not teacher – Set their own goal (plan) – Identify what to learn and choose

material: video lecture, book, (monitor)

– Get feedback on their understanding (manage)

• Human meta-cognition controls

the learning process

S. Suresh, SCE, NTU

Class Room Learning

8/12/2015 3

[1] Wenden, A. L. (1998). Metacognitive knowledge and language learning. Applied Linguistics, 19(4), 515–537. [2] Rivers, W. P. (2001). Autonomy at all costs: An ethnography of meta-cognitive self-assessment and self-management among experienced language learners. Modern Language Journal, 85(2), 279–290. [3] Isaacson, R., & Fujita, F. (2006). Metacognitive knowledge monitoring and self-regulated learning: Academic success and reflections on learning. Journal of the Scholarship of Teaching and Learning, 6(1), 39–55.

Definition of Metacognition

• “The awareness and knowledge of one’s

mental processes such that one can monitor, regulate and direct them to a desired goal”

– As defined by J.H Flavell (1976)


What is self- regulation?

• For effective learning, learners employ self- regulation [2],[3].


[2] Rivers, W. P. (2001). Autonomy at all costs: An ethnography of meta-cognitive self-assessment and self-management among experienced language learners. Modern Language Journal, 85(2), 279–290. [3] Isaacson, R., & Fujita, F. (2006). Metacognitive knowledge monitoring and self-regulated learning: Academic success and reflections on learning. Journal of the Scholarship of Teaching and Learning, 6(1), 39–55.

Definition

• Self-regulation – An active constructive process whereby learners

set goals, monitor, regulate, and control their cognitive and metacognitive process in the service of their goals

– Provide role in collaborative learning


Why Meta-cognition is important? • Learning

– How-to-learn? – What-to-learn? – When-to-learn?

• Helps to promote “Deep Learning” • Assessment for learning

– Active role in assessing his/her own learning – Encourage to take responsibilities – Provide awareness


Models of Metacognition

Nelson and Naren Model • Cognitive component

• Represent the knowledge • Metacognitive component

• Represent dynamic model of the cognitive component

• Signals – Control

• Change the state of cognitive component or cognitive component itself

• Initiate, or terminate or continue

– Monitory • Inform about cognition

Nelson and Narens Model

S. Suresh, SCE, NTU

Nelson T. O and Narens L, Metamemory: A theoretical framework and new findings, Psychology and Learning Motivation, 26, pp. 125-173, 1990.

8/12/2015 8

Part I How metacognition is incorporated

in neural networks?


Metacognitive network

Nelson Naren’s Model Metacognitive network


Meta-cognitive network Cognitive component

• Representation of knowledge – To be learnt from the sample

stream – Unknown

• Suitable structure and its parameters

• Choice of knowledge representation – Neural network : RBFN – Neuro-Fuzzy – Complex-valued neural

network – etc..

Meta-Cognitive component

• Learning about learning – Decides

• What-to-learn – Proper choice of samples

from stream based on current state of knowledge

• When-to-learn – Appropriate usage of

sample in right interval

• How-to-learn – Structure modification – Parameter learning


Current state of metacognitive networks

• Neural Network – G. Sateesh Babu, S. Suresh, Sequential projection based metacognitive learning in a Radial

basis function network for classification problems, IEEE Trans, on Neural Networks and Learning Systems, 24(2), pp. 194-206, 2013.

– G. Sateesh Babu, and S. Suresh, Meta-cognitive RBF networks and Its Projection based Learning Algorithm for Classification problems, Applied Soft Computing, 13(1), pp. 654-666, 2013.

– G. Sateesh Babu, and S. Suresh, Meta-cognitive neural network for classification problems in a sequential learning framework, Neurocomputing, 81(1), pp. 86-96, 2012.

• Neuro-Fuzzy systems – K. Subramanian, S. Suresh, N. Sundararajan, “A meta-cognitive neuro-fuzzy inference system

(McFIS) for sequential classification problems, IEEE Trans. on Fuzzy Systems, 2013 – K. Subraminan, and S. Suresh, A meta-cognitive sequential learning algorithm for neuro-fuzzy

inference system, Applied soft computing, 12(11), 36703-3614, 2012. • Complex-valued neural network

– R. Savitha, S. Suresh, and N. Sundararajan, Metacognitive learning algorithm for a fully complex-valued relaxation network, Neural Networks, 32, pp. 309-318, 2012.

– R. Savitha, S. Suresh, and N. Sundararajan, Meta-cognitive learning in a fully complex-valued radial basis function network, Neural Computation, 24(5), pp. 1297-1328, 2012.


https://dl.dropboxusercontent.com/u/5777899/TNNLS_PBL.pdf


https://dl.dropboxusercontent.com/u/5777899/ASOC_PBLMCRBF.pdf

https://dl.dropboxusercontent.com/u/5777899/ASOC_PBLMCRBF.pdf

https://dl.dropboxusercontent.com/u/5777899/McNN.pdf

https://dl.dropboxusercontent.com/u/5777899/McNN.pdf

https://dl.dropboxusercontent.com/u/5777899/McFIS_TFS.pdf

https://dl.dropboxusercontent.com/u/5777899/McFIS_TFS.pdf

https://dl.dropboxusercontent.com/u/5777899/ASOC_MCFIS_final.ps

https://dl.dropboxusercontent.com/u/5777899/ASOC_MCFIS_final.ps

https://dl.dropboxusercontent.com/u/5777899/McFCRN.pdf

https://dl.dropboxusercontent.com/u/5777899/McFCRN.pdf

https://dl.dropboxusercontent.com/u/5777899/NECO.pdf

https://dl.dropboxusercontent.com/u/5777899/NECO.pdf

META-COGNITIVE RBFN AND ITS SEQUENTIAL LEARNING ALGORITHM

• G. Sateesh Babu, S. Suresh, Sequential projection based metacognitive

learning in a Radial basis function network for classification problems, IEEE Trans, on Neural Networks and Learning Systems, 24(2), pp. 194-206, 2013.






McRBF: Schematic Diagram


McRBF: Cognitive Component

• Input Layer: – m neurons, linear

• Hidden Layer – K neurons, Gaussian

• Output Layer – n neurons, linear

S. Suresh, SCE, NTU

K hidden neurons

Output weights

m input neurons

1

1

m

1

K

1y

ˆnyn output neurons

1x

mx

11w

nKw

8/12/2015 15

McRBF: Meta-cognitive component

• Monitory Signals – Predicted Class Label:

– Posterior Probability:

– Maximum Hinge Error:

– Class Specific Spherical Potential:


McRBF: Meta-cognitive component

• Control signals – Sample Deletion Strategy

• Remove similar samples as that of knowledge stored in the network

– Sample Learning Strategy • Learn the current sample by any of the following way

– Neuron Addition: Add new resource to capture novel knowledge – Neuron Deletion: Delete redundant resource – Parameter Update: Update existing knowledge

– Sample Reserve Strategy • Current sample contain information but I will learn it later


Sequential learning algorithm • Projection Based Learning for a Meta-cognitive

Radial Basis Function Network (PBL-McRBFN): – What is Projection Based Learning?

• Evolving learning algorithm • Classifier based on Hinge loss error function minimization • Based on the best human learning strategy, namely, self-

regulated learning. • Uses past knowledge in learning • Fast learning algorithm:

– Input parameters are initialized through meta-cognition – Output weights are estimated as a solution to a set of linear

equations as a linear programming problem.


Projection Based Learning

• Hinge Loss Error Function

• Weight Minimization • Find optimal W


Projection Based Learning-contd

• Minimum energy point is obtained using

• Solving the above equation, we have:

• Which can be written as:

• In matrix form,


• It can be shown that A inverse exists and

hence


Summary: PBL-McRBFN • Initialization: Set sample 1 (t=1) as the first neuron (K=1) • For samples t = 2, …, N

• Cognition: Compute the output based on the current network • Meta-cognition:

– Monitoring: Compute significance of the sample using maximum hinge-loss error, predicted class label, confidence of classifier, and class-specific spherical potential

– Control: Choose suitable learning strategies » Sample Deletion: Delete samples with insignificant knowledge » Sample Learning

• Neuron addition: Add a neuron (K = K+1) if the sample is very significant. Neuron addition threshold is self-regulated.

• Parameter update: Update the output weight if the sample is significant. Parameter update threshold is self-regulated

» Sample reserve: Reserve the sample. Due to the self-regulatory nature of the thresholds, the sample may be used later.


Performance Evaluation

• Benchmark problems – Classification problems from UCI machine learning

repository

• Applications – Whole brain image based Alzheimer’s disease

detection


Benchmark Problems: UCI machine learning repository


Datasets from UCI [8] Datasets #Features #Classes # Samples Impact Factor

Train Test Train Test

Image Segmentation (IS)

19 7 210 2100 0 0

Liver Disorder (LD)

6 2 200 145 0.17 0.14

Ionosphere (Ion)

34 2 100 251 0.28 0.28


[8] Blake, C., & Merz, C. (1998). UCI repository of machine learning databases. Irvine: Department of Information and Computer Sciences, University of California, Irvine. http://archive.ics.uci.edu/ml/.

For other benchmark problems and results, please refer to: G Sateesh Babu, S Suresh, “Meta-cognitive RBF Network and its Projection Based Learning algorithm for classification problems,” Applied Soft Computing, vol. 13, no. 1, pp. 654–666, 2013.

Results on Benchmark Problems Data PBL-McRBFN SRAN SVM

K Nu Testing K Nu Testing K Testing

ηo ηa ηo

ηa

ηo

ηa

IS 50 89 94.2 94.2 47 113 92.3 92.3 127 91.4 91.4

LD 87 116 73.1 72.6 91 151 66.9 65.8 141 71 70.2

ION 18 58 96.4 96.5 21 86 90.8 91.9 43 91.2 88.5


For results to other benchmark problems and statistical studies, please refer to: G Sateesh Babu, S Suresh, “Meta-cognitive RBF Network and its Projection Based Learning algorithm for classification problems,” Applied Soft Computing, vol. 13, no. 1, pp. 654–666, 2013.

10 fold Cross validation results


Observations

• First time in literature, human metacognition principles are integrated in machine learning framework.

• Self-regulation of cognitive component (RBF network) helps in achieving better generalization.

• Sample reserve strategy • Play vital role in Judgment of Learning • Use self-selected samples for validation of addition of

neurons • One can also use it for preventing drift in sequential learning


Application – Whole Brain MR Imaging based Alzheimer's Disease

Detection


Alzheimer’s disease Main threat to public health

~ 30 million AD patients

worldwide – 3.7 million Indians – 5.3 million Americans

AD - progressive,

degenerative disease that leads to memory loss, poor

judgment and problems in learning

At present, researchers know

of no single cause nor of a cure


Diagnosis of Alzheimer’s Disease

Clinical diagnosis criteria

– National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorder Association Criteria (NINDS-ADRDA) (http://m.medicalcriteria.com/crit/neuro_alzheimer.html)

– Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV) (http://www.dnalc.org/view/2221-DSM-IV-criteria-for-Alzheimer-s-disease.html)

Neuropsychological testing

– Mini-Mental Status Examination (MMSE) (http://www.health.gov.bc.ca/pharmacare/adti/clinician/pdf/ADTI%20SMMSE-GDS%20Reference%20Card.pdf)

– Clinical Dementia Rating (CDR) (http://www.biostat.wustl.edu/adrc/)


http://m.medicalcriteria.com/crit/neuro_alzheimer.html

http://www.dnalc.org/view/2221-DSM-IV-criteria-for-Alzheimer-s-disease.html

http://www.dnalc.org/view/2221-DSM-IV-criteria-for-Alzheimer-s-disease.html

http://www.health.gov.bc.ca/pharmacare/adti/clinician/pdf/ADTI SMMSE-GDS Reference Card.pdf

http://www.health.gov.bc.ca/pharmacare/adti/clinician/pdf/ADTI SMMSE-GDS Reference Card.pdf

http://www.biostat.wustl.edu/adrc/

Neuro imaging in AD

MRI - Magnetic Resonance Imaging

– High spatial resolution – Exceptional soft tissue contrast – Can detect minute abnormalities – Can visualise and measure atrophy rates

Advanced MR techniques

– Diffusion Tensor Imaging - Tissue microstructure – Magnetic resonance spectroscopy - Brain

metabolism – Functional MRI - Neural activity

Early detection of AD from MRI is a promising alternative


Data Sets • Open Access Series of Imaging Studies (OASIS) database (OASIS)

(http://www.oasis-brains.org/) – 100 AD patients, 98 controls – Homogenous Y. Fan et. al, “Integrated feature extraction and selection for neuroimage classification,”

Medical Imaging 2009: Image Processing, vol. 7259, no. 1, p. 72591U, 2009. W. Yang et. al, “ICA-based feature extraction and automatic classification of AD-related

MRI data, ICNC, vol. 3, 2010, pp. 1261-1265.

• Alzheimer's Disease Neuroimaging Initiative (ADNI) (http://adni.loni.ucla.edu/) – 232 AD patients, 200 controls (as of Feb 2012) – Heterogenous W. Yang et. al, “ICA-based automatic classification of magnetic resonance images from

ADNI data,” Part III, LSMS/ICSEE 10, pp. 340-347, 2010.


http://www.oasis-brains.org/

http://adni.loni.ucla.edu/

Feature Extraction using VBM

• Voxel-based Morphometry (VBM) – image analysis technique

(J. Ashburner and K. Friston, “Unified segmentation,” NeuroImage, vol. 26, pp. 839–851, 2005.)

– Identifies regional differences in gray

or white matter between groups of subjects

– Whole-brain analysis - does not require a priori assumptions about region of interest

– Fast and fully automated


MRI image and VBM results

Segmented and Smoothed images

MIP from VBM


Imaging Biomarkers

Selected region Brain region names


observations • Results using PBL-McRBF RFE: • Gender-wise analysis

– Male – INSULA • Emotion and consciousness related problems

– Female – PARAHIPPOCAMPAL GYRUS and EXTRA-NUCLUS regions • Memory encoding and retrieval related problems

• Age-wise analysis

– 60-70 – SUPERIOR TEMPORAL GYRUS • Auditory related problems

– 70-80 - PARAHIPPOCAMPAL GYRUS and EXTRA-NUCLUS regions • Memory encoding and retrieval related problems

– 80 and above – HIPPOCAMPUS, LATERAL VENTICAL, PARAHIPPOCAMPAL GYRUS

• Long-term memory, spatial navigation, memory coding and retrival


Conclusions • For the first time, human metacognition principles are

integrated in a machine learning framework. • Self-regulation of cognitive component (RBF network) helps in

achieving better generalization. • McRBF effectively answer what-to-learn, when-to-learn and

how-to-learn by – Sample deletion strategy – Sample learning strategy

• addition/deletion and update – Sample reserve strategy


Other Applications using meta-cognitive neural networks

• Medical Informatics: – Alzheimer’s disease detection (Sateesh Babu et. al., ICML 2012/IEEE TNNLS

2013). – Parkinson’s disease detection (Sateesh Babu et. al., Expert Systems with

Applications, 2013).

• Video Analytics: – Human Action Recognition (K. Subramanian et. al., International Journal of

Neural Systems, 2012). – Human Emotion Recognition (K. Subramanian et. al., Submitted to IEEE TNNLS,

2013).

• Complex-valued Signal Processing: – Human Action Recognition using complex-valued features (R. V. Babu et. al.,

Neurocomputing, 2012) – QAM Equalization (R. Savitha et. al., Neural computation, 2012/IEEE TNNLS

2013) – Adaptive Beamforming (R. Savitha et. al., Neural computation, 2012/IEEE

TNNLS 2013) 8/12/2015 S. Suresh, SCE, NTU 39

Part II How metacognition is incorporated

in PSO?


What is Optimization?


• The process of finding the conditions that give maximum or minimum of a function.

• The act of obtaining best results under given circumstances.

• The function to be optimized is called the “Objective Function” which consists of – Design Variables: Parameters for defining problem – Constraints: Bound the variables to certain values

• The Objective Function is a real function of n variables ),,,( 21 nxxxf

Formulation of Optimization Problem


( )

( )

1 2

minimize theobjective functionmin ( ), , ,.......,subject toconstraints

( ) 00

n

i

i

f x x x x x

c xc x

=

≥

=

( ) ( )2 21 2

2 21 2

1 2

Example

min 2 1

subject : 02

x x

x xx x

− + − − ≤+ ≤

Find the minimum of the function represented by

Global and local maximum

Local Maximum Global Maximum


fitness

Requirements of Optimization


MODEL Process of identifying Objective function, Variables and

Constraints. Mathematical representation of the problem

ALGORITHM Typically, complex models Requires effective and reliable numerical algorithm No universal optimization algorithm Existing algorithms (not limited to)

1. Gradient Methods 2. Dynamic Programming 3. Non-Linear Programming 4. Evolutionary Algorithm 5. Quasi-Newton Method 6. Particle Swarm Optimization

Particle Swarm Optimization


PSO is a population based procedure to find the best possible solution. (J. Kennedy and R.C. Eberhart. Particle swarm optimization. In Proceedings of IEEE Intl. Conf. on Neural Networks, pages 1942--1948, 1995.) Introduced in 1995 Dr. Eberhart and Dr. Kennedy. Motivated from the behavior of Bird Flock and Fish Schooling. The Scenario and Strategy used by birds can be summarized as: Scenario

Birds searching for food Searching for one piece Only know how far food is

Strategy Follow bird nearest to food


• In PSO each member is called as a ‘Particle’ – randomly initialized within the search space – Having fitness value (Objective Function) – And velocity (Directing Flight) – Flies around the search space.

• Flying is adjusted through – own flying experience (Exploration Process) and – flying experience of other particles (Exploitation

Process).

Particle Swarm Optimization

Working of PSO Algorithm


PSO Update Equations


Update Mechanism


Performed by:

Research Group of Swarm Intelligence at Peking University

Ref: www.cil.pku.edu.cn/resources/pso_and_itsvariants

PSO Animation



x

y

fitness min

max

search space

Initialization: • Particles Randomly initialized in the

search space Position Velocity


x

y

search space

fitness min

max

After Few Iteration: • Particles started moving in the search space Attracted to Local optima Attracted to Global optima


x

y

fitness min

max

search space

After Few Iteration: • Particles moving in the search space More are attracted to Global optima Few are Converging to Local optima


x

y

fitness min

max

search space

After Few more Iteration: • Particles started moving in the search space More particles converge to local optima Particles are attracted to a Local optima


x

y

fitness min

max

search space

• Global Optima Attracts the particles Calls the particles from local optima


x

y

fitness min

max

search space

• Particles have started moving towards Global Optima


x

y

fitness min

max

search space

• Towards the end of Iterations Particles are converging towards Global Optima


x

y

fitness min

max

search space

• Final Iteration Finally the particles converge to a Global Optima


PSO Example

Path followed by a particle for convergence towards Global optimum

Variants of PSO Algorithm


• Discrete PSO ……………… – can handle discrete binary variables

• MINLP PSO………… – can handle both discrete binary and continuous

variables.

• Hybrid PSO…………. – Utilizes basic mechanism of PSO and the natural

selection mechanism, which is usually utilized by EC methods such as GAs.

Application


• PSO Can solve – Multi-objective optimization – Mixed integer programming – Difficult optimization problems

• Application – Complex structural design, – Aircraft wing design – Shape optimization – Finance application – Stock market – Power Transmission Network Expansion Planning

(TNEP)

PSO-Summary


• Mathematical Simplicity • Lesser Computational Efforts • Premature Convergence

– All the particles learn simultaneously from ‘Pbest’ and ‘Gbest’ even if they are far from the global optimum.

• Main research areas in PSO

– Parameter Setting – Deciding the neighborhood – Updating Learning Strategies


Explore human learning principles for better performance improvement.

Self Regulating Particle Swarm Optimization Algorithm


Self Regulating Particle Swarm Optimization Algorithm (SRPSO)


Research in human learning psychology (Nelson et al. 1990) Humans are best planners Self regulation leads towards better planning Enables to decide

What to do, when to do and how to do.

Best planner regulates his learning strategies According to current state of knowledge and Perception on the global knowledge

The proposed algorithm: Self Regulating Particle Swarm Optimization

(SRPSO). (M.R. Tanweer et. al., INS, 2015)

Fig: Effective decision making system for PSO analogous to Nelson and Naren’s model (Nelson et al. 1990)

Strategies Introduced in SRPSO


• Self-Regulated Inertia Weight – Used only for the best particle – Accelerates the exploration process – No self and social cognition – Inspired from the best learner: e.g. Hill Climber

• Self-Perception on Search Directions – Humans believe others based on trust – Trust defines the amount of information to be shared. – Partial social exploitation is used for all the other particles.

Strategies in SRPSO


Self Regulating Inertia Weight: Best Particle Increased Inertia weight Enhanced exploration

Self Perception Strategy: Other Particles Perception based selection of dimensions from global best directions Exploration, Self-exploitation and Partial social-exploitation

Red arrows represent perception on search directions

t t+1

SRPSO Update Equations


Experimental Setup


CEC2005 Benchmark functions (Suganthan et al. 2005) Unimodal (F1-F5) Basic Multimodal (F6-F12) Expanded Multimodal (F13-F14) Hybrid Composition (F15-F25)

Parameter settings Inertia weight (Initial = 1.05, Final = 0.5) Vmax = 0.07*Range Swarm Size = 40

Experiments Functions evaluated for 100 times. Mean, Median and Standard Deviation for 30 Dimension and 50 Dimension.

Analysis of Proposed strategies


Func. Algorithm Median Mean STD.

F4

Basic PSO 1.044E+02 1.544E+02 9.842E+01 SRPSO-w (η=0.5) 9.495E+01 1.028E+02 5.577E+01 SRPSO-w (η=1) 8.079E+01 1.001E+02 6.607E+01 SRPSO-w (η=2) 8.874E+01 1.011E+02 6.046E+01

SRPSO-p (λ=0.25) 1.339E+02 1.393E+02 6.071E+01 SRPSO-p (λ=0.5) 5.685E+01 6.889E+01 4.094E+01

SRPSO-p (λ=0.75) 4.560E+01 6.191E+01 4.891E+01 SRPSO (η=1 & λ=0.5) 1.998E+01 2.988E+01 2.480E+01

Individual Strategy and Combined Effect:

Convergence: • SRPSO has significantly improved the performance of PSO.

• Similar observations are made for all the CEC2005 benchmark functions.

• SRPSO-w: SRPSO only with self-regulated inertia weight

• SRPSO-p: SRPSO only with self-perception

• η is taken as ‘1’ because it provides best solution.

• λ is selected as 0.5 because • Lower value will make SRPSO like basic PSO. • Higher value will make particles fly in their own

directions.

Function:

• F4 from CEC2005 • Shifted Schwefel’s

Problem with noise

• Unimodal

0 1 2 3x 105

101

Function Evaluations

Log(

F(x)

-F(x

*))

Normal PSOSRPSO

Selected PSO variants for Comparison


• Clerc et al. “The particle swarm - Explosion, stability, and convergence” (2002). (χPSO) – Introduced Constriction factor (χ) – Better convergence on selected problems (Bui et al., 2010)

• Mendes et al. “The fully informed particle swarm: simpler, maybe better” (2004). (FIPS) – New information flow scheme – Better convergence on multimodal functions (Chen et al., 2013)

• Liang et al. “Dynamic multi-swarm particle swarm optimizer” (2005). (DMS-PSO) – Dynamically changing neighborhood – Slower convergence speed (Nasir et al., 2012)

• Kennedy. “Bare bones particle swarms” (2003). (BBPSO) – Gaussian search space – Better performance on Hybrid Composition functions (Blackwell, 2012)

• Parsopoulos. “On the computation of all global minimizers through PSO” (2004). (UPSO) – Combined effect of global and local variants – Better convergence on selected problems (Epitropakis et al., 2012)

• Liang et al. “Comprehensive learning PSO for global optimization of multimodal” (2006). (CLPSO)

– Particles learn from self-cognition – Better convergence on complex multimodal functions (Nasir et al., 2012)

Results (Unimodal)


Function Algorith

m 30 Dimensions 50 Dimensions

Median Mean STD. Median Mean STD.

F1

χPSO 5.328E+00 9.657E+00 1.233E+01 5.328E+00 9.657E+00 1.233E+01 BBPSO 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00

DMSPSO 1.143E+02 3.135E+02 4.149E+02 2.349E+02 3.870E+02 3.855E+02 FIPS 3.185E+02 5.252E+02 5.571E+02 1.149E+03 1.673E+03 1.524E+03

UPSO 1.269E+03 1.306E+03 7.328E+02 6.840E+02 7.100E+02 3.290E+02 CLPSO 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00 SRPSO 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00 0.000E+00

F2

χPSO 0.000E+00 1.573E+01 8.112E+01 2.334E+02 7.774E+02 1.806E+03 BBPSO 6.000E-03 9.260E-03 8.480E-03 2.407E+02 2.886E+02 1.452E+02


UPSO 6.688E+03 7.602E+03 5.290E+03 3.632E+03 4.220E+03 2.894E+03 CLPSO 3.828E+02 3.828E+02 1.060E+02 1.013E+04 1.021E+04 1.357E+03 SRPSO 0.000E+00 0.000E+00 0.000E+00 1.400E-03 3.767E-03 1.753E-02

F3




3 Algorithms have converged to the optimum solution

SRPSO has converged to the optimum solution

SRPSO has provided much better results than the other algorithms

Results (Multimodal)


Func. Algorithm 30 Dimensions 50 Dimensions


F6




F7

χPSO 6.788E+03 6.780E+03 1.291E+02 6.158E+03 6.154E+03 7.416E+01 BBPSO 4.696E+03 4.696E+03 5.039E-01 6.195E+03 6.197E+03 4.553E+00


UPSO 7.513E+03 7.524E+03 3.409E+02 7.419E+03 7.420E+03 3.034E+02 CLPSO 4.696E+03 4.696E+03 1.837E-12 6.195E+03 6.195E+03 4.594E-12 SRPSO 2.940E-02 7.610E-02 9.962E-02 5.569E-01 5.444E-01 2.253E-01

F12




Best Solution in 50 Dimension

Better Solutions several order better than all other variants

Better mean performance in both dimension.

Results (Hybrid Composition)


Func. Algorith

m 30 Dimensions 50 Dimensions


F16




F19




F25




Better median performance

Better performance in both dimension

Better mean performance in both dimension.

Results Analysis


Better solutions in both 30D and 50D cases. Best solutions All unimodal functions.

4 out of 7 basic multimodal functions.

1 of the 2 expanded multimodal functions.

8 out of 11 Hybrid composition functions.

Summary


PSO is a simple and effective optimization algorithm. It experiences premature convergence. It has been extensively researched. Incorporating human learning principles in PSO is a new

search direction. SRPSO is human self-learning inspired PSO variant. SRPSO has significantly enhanced PSO convergence

characteristics. Faster convergence closer to optimum solution has

been observed.

Limitations in SRPSO

S. Suresh, SCE, NTU 78

• Only incorporates Human self-cognition

• Same perception for all particles

• Diversity management is not proper

• Performance suffers on few functions

• Need of addressing Human social behaviour 8/12/2015

Mentoring based Particle Swarm Optimization Algorithm


Motivation


Human social learning in SLPSO (Cheng & Jin, 2015)

Human self-cognition in SRPSO (Tanweer et al., 2015)

Need to address both self and social learning together

Mentoring based learning Process of positive learning within a group. Dynamic Learning environment Effective Learners act as Mentor for less efficient learners, the Mentees Moderate Learners perform independent learning Performance decides the learners

8/12/2015

Mentoring based Particle Swarm Optimization (MePSO) Algorithm

S. Suresh, SCE, NTU 81 8/12/2015

Learning Strategies


Mentors Group Higher self-cognition and partial social cognition Best Particle: Self-regulating inertia weight from

SRPSO (Tanweer et al., 2015)

Mentee Group Either mentor or self guidance

Independent Learners Group Self-perception strategy (Tanweer et al., 2015)

8/12/2015

Convergence Analysis: Impact of Mentoring


0 1000 2000 3000 4000 5000

101

Iterations

Log(

F(x)

-F(x

*))

PSOSRPSOMePSO

8/12/2015

Real World Application

8/12/2015 84 S. Suresh, SCE, NTU

Problem Definition

85

Transmission Network Expansion Planning (TNEP) problem (I de J Silva et. al., IET Proceedings- Generation Transmission and Distribution, Dec 2005)

Determine the set of new lines to be constructed Cost of Expansion is minimum No overload

The Problem set includes: Generating Points, generating capacity and voltage level Load point and load value Existing lines and transformer units Investment cost of lines, power rating and transformer Power losses cost

The dynamic formulation is a large scale non-linear mixed interger optimization problem.

8/12/2015 S. Suresh, SCE, NTU

Problem Formulation

86 8/12/2015 S. Suresh, SCE, NTU

Problem Formulation

87 8/12/2015 S. Suresh, SCE, NTU

EXAMPLE



89

• Evaluated following the guidelines of CEC2011 – 25 independent run – Swarm size = 50 – Compared with top 2 best performing algorithms

• GA-MPC - Genetic Algorithm with a new Multi-Parent Crossover: An efficient and improved variant of GA with a new crossover operator. The algorithm has produced robust and high quality solution to optimization problems.

• SAMODE – Self-Adaptive Multi-Operator Differential Operator: A new variant of DE with four different mutation types and one crossover operator with each operator assigned a sub-population. The algorithm has successfully addressed problems with diverse classes.



90

GA-MPC SAMODE MePSO

2.200E+02 2.200E+02 2.200E+02

All the algorithms have produced same results BUT

1. MePSO has achieved the solution within 50% of function evaluations. 2. MePSO has a swarm size of 50 compared to 100 of other two algorithms 3. MePSO is computationally efficeint for real-world problems.


Publications


• Tanweer, M. R., S. Suresh, and N. Sundararajan. "Self regulating particle swarm optimization algorithm“. Information Sciences 294 (2015): 182-202.

• Tanweer, M. R., S. Suresh, and N. Sundararajan. Dynamic Mentoring and Self-

Regulation based Particle Swarm Optimization Algorithm for solving Complex Real-world Optimization Problems. Information Sciences. 326 (2016): 1-24.

• Tanweer, M. R., Suresh, S., & Sundararajan, N. (2015, May). Mentoring based particle swarm optimization algorithm. Evolutionary Computation (CEC), 2015 IEEE Congress on. IEEE, 2015.

• Tanweer, M. R., Suresh, S., & Sundararajan, N. (2015, May). Improved SRPSO algorithm for solving computationally expensive numerical optimization problems. Evolutionary Computation (CEC), 2015 IEEE Congress on. IEEE, 2015.

Additional Info.


• Papers can be downloaded from ResearchGate https://www.researchgate.net/profile/Muhamm

ad_Tanweer • For Software: Please contact M.R. Tanweer at [email protected]

https://www.researchgate.net/profile/Muhammad_Tanweer

https://www.researchgate.net/profile/Muhammad_Tanweer

mailto:[email protected]

Some Open problems • Issue in McRBF: Current framework is a static implementation of

metacognition – Fixed control signals – Monitory signals are based on current samples. They do not reflect

feeling-of-knowing, judgment-of-knowledge and ease-of-learning • Human Thinking

– Common sense influence learning significantly – Introspective/retrospective thinking influence the learning

significantly. They are responsible for dynamic change in control from meta-cognition

• Social Learning – Current framework does not consider the cooperation/collaboration in

learning • Transfer Knowledge – one domain to other


Diagnosis of Alzheimer’s Disease (contd)

Neuroimaging: • Computed Tomography (CT): O. L. Lopez et. al, Computed tomography

but not magnetic resonance imaging identified periventricular white-matter lesions predict symptomatic cerebrovascular disease in probable Alzheimer's disease, Archives of Neurology, vol. 52, pp. 659-664, 1995.

• Single-Photon Emission Computed Tomography (SPECT): J. Ramrez et. al, Early detection of the Alzheimer's disease combining feature selection and kernel machines. Advances in Neuro-Information Processing, vol. 5507, pp. 410-417, 2009.

• Positron Emission Tomography (PET): M. Lopez et.al, Principal component analysis based techniques and supervised classification schemes for the early detection of Alzheimer's disease, Neurocomputing, vol. 74, pp. 1260-1271, 2011.

• Magnetic Resonance Imaging (MRI): S. Kloppel et.al, Automatic classfication of MR scans in Alzheimer's disease, Brain, vol. 131, pp. 681-689, 2008.


human metacognition inspired learning algorithms in neural

Documents